On 1/24/21 3:30 PM, Ard Biesheuvel wrote:
> On Sun, 24 Jan 2021 at 15:28, Helge Deller <deller@xxxxxx> wrote:
>>
>> On some of my parisc machines, this patch c055e3eae0f1 ("crypto: xor -
>> use ktime for template benchmarking") triggers a dividy-by-zero
>> exception because "min" becomes zero, which then leads to a kernel
>> crash.
>>
>> It's not clear yet, why I see the issue only on some machines. At least
>> on those the measured time to run the xor speed tests becomes zero
>> nanoseconds. Maybe it's because gcc-10 optimizes the speed test out, or
>> because of some other changes in the time keeping routines.
>>
>> In either case, the kernel should not crash.
>>
>> This patch adds a workaround by reporting such cases with a kernel
>> warning and continues as if the xor tests would have run in 1 ns.
>>
>> Signed-off-by: Helge Deller <deller@xxxxxx>
>> Cc: stable@xxxxxxxxxxxxxxx # 5.10+
>>
>> ---
>>
>> diff --git a/crypto/xor.c b/crypto/xor.c
>> index eacbf4f93990..3639341bac7e 100644
>> --- a/crypto/xor.c
>> +++ b/crypto/xor.c
>> @@ -100,6 +100,8 @@ do_xor_speed(struct xor_block_template *tmpl, void *b1, void *b2)
>> mb();
>> }
>> diff = ktime_sub(ktime_get(), start);
>> + if (WARN_ON(diff == 0))
>> + diff = 1;
>> if (diff < min)
>> min = diff;
>> }
>
> This should already be fixed in mainline - please check whether that
> fix works for you.
Yes, it's basically the same fix and a good step.
It fixes the kernel crash at least, but it's still strange that
the timing calculation will be wrong.
Helge
